Learning in a noisy environment: Adaptive management for inconvenient models

• Environmental and Resource Economics Workshop

Date and Location

Friday, October 1, 2010, 12:10 PM - 1:00 PM

Abstract

In this paper we extend the deterministic mangrove fishery, ecosystem service model of Sanchirico and Springborn (In press) to incorporate both risk (irreducible uncertainty) and uncertainty that is reduced over time through learning. We demonstrate how to handle learning models that are more realistic but "inconvenient" in the sense that the function describing a decision-maker's updated beliefs about the nature of the system does not follow a convenient closed (conjugate) form. To facilitate this we develop a method for approximating the decision-maker's posterior beliefs using a Kullback–Leibler divergence approach. The full management model describes optimal management of both a harvestable resource (fish) and the ecosystem on which it may depend (mangroves) as a function of the state of the system, which includes both fish and mangrove stocks as well as current information or beliefs about mechanics. Since describing a non-trivial information state space will typically involve at least two state variables, overall we solve a four-state, two-control dynamic programming problem. Since the rate of learning over time is influenced endogenously by both control variables our setting is one of active adaptive management where current controls are selected to maximize returns which include the expected value of information.